Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。

01.
arXiv (CS.LG) 2026-06-15

Beyond a Single Explanation of the Adam–SGD Gap

arXiv:2606.14259v1 Announce Type: new Abstract: Prior work has identified several factors that can contribute to the performance gap between Adam and SGD, spanning data aspects, architecture design, and optimization properties. Yet these explanations are often studied in isolation, leaving their relative importance unclear. In this work, we revisit these hypotheses through a controlled empirical study across vision, language, genomics, and graph tasks, spanning modern and classical architectures, and carefully designed training setups. Our results suggest that no single factor consistently explains the Adam–SGD gap. For instance, the Adam advantage can (1) persist under a uniform vocabulary distribution yet nearly disappear under a heavy-tailed one; (2) reverse in favor of SGD in softmax-attention models; and (3) become larger under soft architectural modifications, e.g., when ReLU is replaced by a GeLU nonlinearity. This suggests that the gap arises from nontrivial data and architecture interactions, rather than from a single common factor. Yet, we observe a pattern across our settings: a crossover batch size at which the relative advantage shifts from SGD to Adam as the batch size scales. These empirical results are captured by our theoretical gap model, which predicts this batch-size-dependent crossover. Our perspective helps reconcile several existing hypotheses while offering practical insights across domains.

02.
arXiv (CS.CV) 2026-06-16

Double-Helix Vision (DH-V2): A Geometry-Based Visual Sampler for Bandwidth-Constrained Perception

作者:

We present Double-Helix Vision (DH), a geometry-based visual sampler that compresses 2D images into compact 1D signals using paired golden-ratio-inspired spiral trajectories. Rather than processing every pixel uniformly, DH employs two phase-shifted helices (Alpha and Beta, offset by 180 degrees) to sample the image with biologically-inspired foveation: high density at the center, sparse coverage at the periphery. At 4K resolution, DH achieves a 1,433x compression ratio (99.93% reduction) while preserving the geometric structure of the scene. The full perception pipeline – including spatial mapping, temporal collision detection, and intra-frame structural disparity estimation – runs in 0.52 ms at 1080p on CPU-only hardware, with no neural network dependencies. On CIFAR-10 at extreme sampling budgets (K=128 points per helix), DH achieves a +6.03% accuracy gain over uniform random sampling. A JSON-serializable Robotics API is provided, delivering sub-millisecond spatial perception reports in 2.7 KB packets. Code and benchmarks are available under the MIT License.

03.
arXiv (CS.AI) 2026-06-16

NeuroSymbolic AI for Legal AI-TRISM: Trustworthy, Reliable, Interpretable, Safe Models

arXiv:2606.15646v1 Announce Type: new Abstract: Large Language Models (LLMs) have transformed natural language processing, but their lack of interpretable reasoning and tendency to hallucinate pose significant challenges for legal applications. While LLMs show promise for legal text analysis and generation, they struggle with accurate citation attribution and precedent verification. For example, in legal contexts, a single incorrect precedent can jeopardize a case. Current approaches to improve LLM reliability in legal domains suffer from two key limitations: inadequate integration of structured legal knowledge during training or fine-tuning, and insufficient verification mechanisms for generated legal content. To address these challenges, we propose the TRISM (Trustworthy, Reliable, Interpretable, Safe Models) framework, which integrates NeuroSymbolic AI principles with LLMs to leverage both neural learning capabilities and symbolic reasoning over structured legal knowledge. The TRISM approach addresses the above limitations while maintaining interpretable decision pathways. Our framework formalizes the extraction of symbolic knowledge from legal textual documents and incorporates Retrieval-Augmented Generation (RAG) as a core component for grounding LLM outputs in verified legal sources. In this position paper, we make the following contributions: (1) An analysis of the limitations of AI in law; (2) Introduce RASOR RAG which creates foundations for neurosymbolic RAG by generating explicit interpretable rationales that could be formalized into symbolic representations; (3) A formalized methodology for creating symbolic legal knowledge bases that support both interpretable reasoning and output verification in LLMs; and (4) The TRISM framework for integrating symbolic legal knowledge with LLMs.

04.
arXiv (CS.CL) 2026-06-16

Tying the Loop – Tied Expert Layers in Mixture-of-Experts Language Models

作者:

Mixture-of-Experts (MoE) architectures efficiently scale Large Language Models (LLMs) by activating only a small fraction of their experts per token, yet the full parameter count - dominated by the expert parameters - must be held in training and inference memory. To address this, we introduce Expert Tying, an architectural modification that shares expert parameters across consecutive transformer layers while preserving independent, layer-wise routing and attention. We evaluate this approach across common, state-of-the-art architectures, including OLMoE, Qwen3, and DeepSeek-style MoEs. Our pretraining experiments demonstrate that tying experts can reduce memory footprint by almost 2x at virtually no degradation in perplexity or downstream quality. By exploiting the parameter redundancy inherent in MoE pathways, our method provides a highly favorable compute-to-memory trade-off, advancing efficient training and scaling of next-generation LLMs.

05.
arXiv (CS.CV) 2026-06-16

GroupToM-Bench: Benchmarking Group Theory of Mind and Nonlinear Social Emergence in MLLMs

True general intelligence requires not only a model of the physical world but also a social world model: the capacity to infer how individual mental states interact and crystallize into group-level outcomes. Despite notable progress in individual-level Theory of Mind (ToM) reasoning, existing multimodal large language models fail at this broader task. Collective behavior emerges non-linearly from social tensions, conformity dynamics, and structural constraints, meaning it cannot be recovered by merely summing individual intentions. We present GroupToM-Bench, the first multimodal benchmark for group-level ToM, built around a causal chain spanning micro-level BDI states (belief, desire, intention), meso-level group tension and structural constraints, and macro-level outcome prediction and mechanistic attribution. To probe this full arc, we develop a seven-level cognitive audit framework. Experiments reveal a gap between current models and human baselines, highlighting a failure to process social structures and non-linear collective dynamics.

07.
arXiv (CS.LG) 2026-06-12

Multimodal Graph Negative Learning

arXiv:2606.12863v1 Announce Type: new Abstract: Multimodal attributed graphs (MAGs) integrate graph topology with heterogeneous modality attributes, such as text and images, thereby enabling richer modeling of complex relational systems. However, such expressiveness also makes learning on MAGs depend on multiple semantic sources, including structural topology, textual and visual attributes, each of which can be regarded as a branch for node representation. Node-level branch semantic imbalance arises when these branches differ across nodes in semantic informativeness and reliability: a branch that provides discriminative semantics for one node may mislead another due to bias in modality quality or structural context. Existing methods often mitigate such heterogeneity through cross-branch agreement or alignment, implicitly treating the dominant prediction as reliable supervision. When the dominant branch is biased, forced imitation may propagate its bias to other branches and suppress original semantics that are useful for classification. We propose GraphMNL, a graph-aware multimodal negative learning framework that addresses this issue by using Negative Learning as cross-branch guidance. Instead of forcing inferior branches to imitate a teacher prediction, the model teaches them which classes a node is unlikely to belong to. GraphMNL builds a branch library, identifies dominant and inferior branches via graph-aware reliability arbitration, gates unstable transfer, and applies target-preserving negative learning over non-target classes. This design decouples target supervision from branch guidance so that supervised losses learn the correct class, while Negative Learning suppresses unlikely alternatives when branch agreement is unreliable. Through the comprehensive experimental evaluation, GraphMNL achieves the best performance on Grocery datasets with 72.47% accuracy and 76.60 F1 score on Reddit M datasets.

08.
medRxiv (Medicine) 2026-06-10

Developmental Associations Linking Childhood Trauma and Early Cannabis Use to Adolescent DNA Methylation and Psychotic-Like Experiences

Background. Psychotic-like experiences (PLEs) index early risk for psychotic disorders and are consistently associated with childhood trauma, yet underlying biological mechanisms remain poorly understood. DNA methylation (DNAm) may capture the biological embedding of early adversity, while adolescent exposures such as cannabis use may modify these processes. We examined epigenome-wide associations of childhood trauma and PLEs, tested the moderating role of early cannabis use, and evaluated DNAm as a potential mediator. Methods. We analysed data from the Avon Longitudinal Study of Parents and Children (ALSPAC), a UK population-based birth cohort. Childhood trauma was assessed prospectively and retrospectively. Epigenome-wide DNAm was measured in peripheral blood at ~17 years using the Illumina 450K array, and PLEs were assessed at 18 using a structured interview. Epigenome-wide association studies were conducted for trauma-DNAm and DNAm-PLEs associations in the final sample (n = 1,457), adjusting for demographic, biological, and technical covariates. Differentially methylated regions (DMRs) were identified using DMRff, followed by functional enrichment analyses. Cannabis use at 15.5 was modelled as a moderator with multiple imputation for missing data. Mediation was tested using the Divide-Aggregate Composite-null Test (DACT). Results. Childhood trauma was associated with widespread DNAm differences, primarily at the regional level, with enrichment in pathways related to cellular stress responses. In contrast, DNAm associated with PLEs was more limited and implicated loci involved in epigenetic regulatory processes. These signatures were largely distinct, and there was no evidence supporting mediation after multiple testing correction. Incorporating cannabis use altered the pattern and extent of DNAm associations, with stronger and more significant signals observed at both CpG and regional levels, although these did not translate into evidence of mediation. Conclusion. Childhood trauma and PLEs show distinct DNAm signatures in adolescence, with trauma-related DNAm reflecting broad stress-related processes and PLE-associated DNAm implicating regulatory mechanisms. We found little evidence that DNAm mediates the trauma-PLE association. Instead, adolescent exposures, particularly cannabis use, may distinctly influence trauma-related epigenetic variation with limited detectable downstream effects on PLEs. These findings support a context-dependent model of epigenetic risk and highlight the need for larger longitudinal studies to clarify causal pathways linking early adversity to psychosis.

09.
arXiv (CS.CV) 2026-06-17

Mordal: Automated Pretrained Model Selection for Vision Language Models

Incorporating multiple modalities into large language models (LLMs) is a powerful way to enhance their understanding of non-textual data, enabling them to perform multimodal tasks. Vision language models (VLMs) form the fastest growing category of multimodal models because of their many practical use cases, including in healthcare, robotics, and accessibility. Unfortunately, even though different VLMs in the literature demonstrate impressive visual capabilities in different benchmarks, they are handcrafted by human experts; there is no automated framework to create task-specific multimodal models. We introduce Mordal, an automated multimodal model search framework that efficiently finds the best VLM for a user-defined task without manual intervention. Mordal achieves this both by reducing the number of candidates to consider during the search process and by minimizing the time required to evaluate each remaining candidate. Our evaluation shows that Mordal can find the best VLM for a given problem using $8.9\times$–$11.6\times$ lower GPU hours than grid search. We have also discovered that Mordal achieves about 69\% higher weighted Kendall's $\tau$ on average than the state-of-the-art model selection method across diverse tasks.

10.
arXiv (CS.CV) 2026-06-15

Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models

Text-to-image (T2I) systems increasingly rely on Large Language Model (LLM)-based text conditioning to interpret and expand user prompts. While this improves prompt understanding and text-image alignment, we find that it can also introduce implicit demographic assumptions, even when demographic attributes are unspecified. To systematically investigate this behavior across varying levels of prompt ambiguity and complexity, we construct a comprehensive benchmark covering diverse prompt settings. Evaluations on eight recent T2I models show that LLM-based systems consistently exhibit stronger demographic skew than non-LLM-based baselines. We further analyze system prompts, a component unique to LLM-based T2I systems that guides prompt interpretation and expansion. Our analyses show that these instructions strongly influence text embeddings, which subsequently leads to biased image generations. Motivated by these findings, we propose FairPro, a training-free debiasing framework that adaptively generates fairness-aware instructions while preserving user intent. Experiments demonstrate that FairPro substantially reduces demographic disparities while maintaining prompt fidelity.

11.
arXiv (CS.CL) 2026-06-16

Not All Skills Help: Measuring and Repairing Agent Knowledge

LLM agents can improve without weight updates by accumulating natural-language skills from experience, but current systems entrust every decision about which skills to keep and how to apply them to LLM judgment alone. We argue that this conflates two distinct roles: generating a skill from experience is a creative act that judgment handles well, while deciding whether that skill actually helps requires empirical evidence across many tasks. Measuring per-skill causal contributions via randomized masking, we find that skill libraries exhibit pervasive causal heterogeneity: individual skills routinely help on some task types while hurting on others, yet their opposing effects cancel in aggregate, making them invisible to global curation methods. We propose ASSAY, a framework that separates generation from curation: it computes a per-skill causal attribution on a small development set, restructures the library offline, and suppresses skills with negative predicted effect for each test task. Across seven base models spanning four providers and two benchmarks (AppWorld and tau-bench), ASSAY consistently improves over prior skill-curation approaches. On AppWorld's hardest split, DeepSeek-V3 achieves 69.3% task-goal completion (47.4% relative improvement), a new state of the art among all published methods including weight-tuned approaches. On tau-bench retail, GPT-4.1 improves by 8.7% relative, advancing past o4-mini, o1, and GPT-4.5 on the public leaderboard without any weight modification. Ablation traces the dominant gain to per-task masking, confirming that the bottleneck is matching skills to tasks at inference time, not removing bad skills globally. Code is available at https://github.com/aiming-lab/assay.

12.
arXiv (CS.AI) 2026-06-17

Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement

arXiv:2606.18247v1 Announce Type: cross Abstract: Robots deployed in the real world should learn from their experience and improve over time. This requires a mechanism of practicing and learning from feedback. In this paper, we propose VERITAS, a generator-verifier framework for generalist robot policies for inference-time policy steering and self-improvement. We use a pre-trained generalist robot policy as a ``generator'' and pair it with a gradient-free ``visual verifier'' that evaluates actions at inference time. This framework enables inference-time steering that improves policy performance without additional training. We demonstrate that inference-time verification consistently outperforms vanilla generalists without training on additional demonstration data. Additionally, we demonstrate that the verified rollouts provide effective supervision for offline policy improvement: policies fine-tuned on verified self-generated trajectories achieve consistent performance gains. Notably, we find that post-training with verified rollouts achieves comparable efficiency to expert demonstrations, while requiring no human interventions. Our results highlight inference-time verification as a practical and scalable mechanism for improving robotic policies during deployment.

13.
arXiv (CS.CV) 2026-06-16

HMR-Net: Hierarchical Modular Routing for Cross-Domain Object Detection in Aerial Images

Despite advances in object detection, aerial imagery remains a challenging domain, as models often fail to generalize across variations in spatial resolution, scene composition, and semantic label coverage. Differences in geographic context, sensor characteristics, and object distributions across datasets limit the capacity of conventional models to learn consistent and transferable representations. Shared methods trained on such data tend to impose a unified representation across fundamentally different domains, resulting in poor performance on region-specific content and less flexibility when dealing with novel object categories. To address this, we propose a novel modular learning framework that enables structured specialization in aerial detection. Our method introduces a hierarchical routing mechanism with two levels of modularity: a domain routing layer that uses latent geographic embeddings to assign inputs to domain-specialized expert modules, and a scene routing mechanism that allocates image subregions to scene-specific expert modules. This allows our method to specialize across datasets and within complex scenes. Additionally, the framework contains a conditional expert module that uses external semantic information (e.g., category names or textual descriptions) to enable detection of novel object categories during inference, without the need for retraining or fine-tuning. By moving beyond monolithic representations, our method provides an adaptive framework for remote sensing object detection. Comprehensive evaluations on four datasets highlight improvements in multi-dataset generalization, region-level specialization, and open-category detection.

14.
arXiv (CS.CV) 2026-06-17

GOT-JEPA: Generic Object Tracking with Model Adaptation and Occlusion Handling using Joint-Embedding Predictive Architecture

The human visual system tracks objects by integrating current observations with previously observed information, adapting to target and scene changes, and reasoning about occlusion at fine granularity. In contrast, recent generic object trackers are often optimized for training targets, which limits robustness and generalization in unseen scenarios, and their occlusion reasoning remains coarse, lacking detailed modeling of occlusion patterns. To address these limitations in generalization and occlusion perception, we propose GOT-JEPA, a model-predictive pretraining framework that extends JEPA from predicting image features to predicting tracking models. Given identical historical information, a teacher predictor generates pseudo-tracking models from a clean current frame, and a student predictor learns to predict the same pseudo-tracking models from a corrupted version of the current frame. This design provides stable pseudo supervision and explicitly trains the predictor to produce reliable tracking models under occlusions, distractors, and other adverse observations, improving generalization to dynamic environments. Building on GOT-JEPA, we further propose OccuSolver to enhance occlusion perception for object tracking. OccuSolver adapts a point-centric point tracker for object-aware visibility estimation and detailed occlusion-pattern capture. Conditioned on object priors iteratively generated by the tracker, OccuSolver incrementally refines visibility states, strengthens occlusion handling, and produces higher-quality reference labels that progressively improve subsequent model predictions. Extensive evaluations on seven benchmarks show that our method effectively enhances tracker generalization and robustness.

15.
arXiv (CS.AI) 2026-06-11

An XAI View on Explainable ASP: Methods, Systems, and Perspectives

arXiv:2601.14764v2 Announce Type: replace Abstract: Answer Set Programming (ASP) is a popular declarative reasoning and problem solving approach in symbolic AI. Its rule-based formalism makes it inherently attractive for explainable and interpretive reasoning, which is gaining importance with the surge of Explainable AI (XAI). A number of explanation approaches and tools for ASP have been developed, which often tackle specific explanatory settings and may not cover all scenarios that ASP users encounter. In this survey, we provide, guided by an XAI perspective, an overview of types of ASP explanations in connection with user questions for explanation, and describe their coverage by current theory and tools. Furthermore, we pinpoint gaps in existing ASP explanations approaches and identify research directions for future work.

16.
arXiv (CS.AI) 2026-06-16

Visualizing Uncertainty: Spatial Maps of Missing and Conflicting Evidence in Deep Learning

arXiv:2606.15767v1 Announce Type: cross Abstract: Understanding when and why deep neural networks are uncertain is crucial for deploying reliable machine learning systems in safety-critical domains. While existing uncertainty quantification methods provide scalar measures of model confidence, they offer limited insight into which spatial regions of an input contribute to different types of uncertainty. We propose a novel visualization framework, Uncertainty Activation Map (UAM), that combines Evidential Deep Learning (EDL) with Full-Gradient Class Activation Mapping (FullGrad) to generate interpretable spatial uncertainty activation maps. Our approach distinguishes between two fundamental types of uncertainty: vacuity, representing lack of evidence, and dissonance, capturing conflicting evidence between competing hypotheses. By leveraging the complete gradient decomposition property of FullGrad and the principled uncertainty quantification of Subjective Logic, our method produces theoretically grounded visualizations that highlight specific image regions responsible for model uncertainty. With this framework, vacuity and dissonance activation maps are generated by computing belief-weighted attributions, enabling identification of where models lack knowledge versus where they encounter ambiguous evidence. Extensive evaluations across multiple benchmark datasets demonstrate that the proposed framework effectively addresses the critical gap between uncertainty quantification and explainability, providing intuitive visual feedback to assess model reliability in complex visual recognition tasks.

17.
bioRxiv (Bioinfo) 2026-06-16

Expanding gene regulatory networks from transcriptome data through graphical modeling with heterogeneous priors

Gene regulatory network inference is widely used to reconstruct large-scale networks and identify functional genes from transcriptome data. Meanwhile, in many biological fields, core regulatory genes have been extensively studied, leading to the establishment of small-scale gene regulatory networks, and novel genes connected to these networks remain to be identified. However, methods for expanding existing gene networks by identifying novel regulatory interactions, rather than reconstructing the entire network, are not well established. Here, we propose a method for gene network expansion that incorporates known regulatory relationships and evaluates each candidate gene individually to infer its regulatory connections to the existing network. Using simulated datasets from the DREAM4 benchmark and the PRECISE-1K experimental dataset, our method outperformed conventional methods by incorporating prior knowledge. In particular, it improved the ability to distinguish true regulatory interactions from indirect associations arising from strong correlations among genes in the existing network. The method also showed strong performance for interactions involving genes with high outdegree or centrality. Furthermore, it maintained stable performance as the size of the existing network increased and was robust to noise in prior information. These results demonstrate that our method provides an effective framework for expanding existing gene regulatory networks by leveraging prior knowledge.

18.
arXiv (CS.AI) 2026-06-16

AlignCoder: Aligning Retrieval with Target Intent for Repository-Level Code Completion

arXiv:2601.19697v2 Announce Type: replace-cross Abstract: Repository-level code completion remains a challenging task for existing code large language models (code LLMs) due to their limited understanding of repository-specific context and domain knowledge. While retrieval-augmented generation (RAG) approaches have shown promise by retrieving relevant code snippets as cross-file context, they suffer from two fundamental problems: misalignment between the query and the target code in the retrieval process, and the inability of existing retrieval methods to effectively utilize the inference information. To address these challenges, we propose AlignCoder, a repository-level code completion framework that introduces a query enhancement mechanism and a reinforcement learning based retriever training method. Our approach generates multiple candidate completions to construct an enhanced query that bridges the semantic gap between the initial query and the target code. Additionally, we employ reinforcement learning to train an AlignRetriever that learns to leverage inference information in the enhanced query for more accurate retrieval. We evaluate AlignCoder on two widely-used benchmarks (CrossCodeEval and RepoEval) across five backbone code LLMs, demonstrating an 18.1% improvement in EM score compared to baselines on the CrossCodeEval benchmark. The results show that our framework achieves superior performance and exhibits high generalizability across various code LLMs and programming languages.

19.
Nature (Science) 2026-06-17

Mapping the neuronal building blocks of human language with language models

作者:

Humans can convey new and highly diverse information through language. This ability to form and combine words into elaborate phrases and sentences enables us to express inexhaustible meanings and is fundamental to human cognition1–5. However, understanding the microscopic cellular building blocks and cortical landscape that precisely underlie human language has remained a challenge. Here we used wide-scale single-neuronal recordings combined with natural language processing models to identify fine-grained linguistic representations across the human frontotemporal cortex during language production. We find that, whereas certain neurons represented the detailed grammatical relationships between words or their parts of speech, others tracked the sentences’ higher-order syntactic structure, their phrase transitions and sequence. Collectively, these neurons reliably captured the words’ syntactic and semantic properties but also dynamically incorporated their specific sentence contexts, therefore enabling them to encode information combinatorially and at highly granular levels of detail. We show how these cell populations were locally organized and how their microscale representations differed from that of their wider field potential patterns. We also show how these neurons were distributed broadly across the frontotemporal cortex, but how their ability to encode linguistic information was left-lateralized and varied between cortical regions. Together, these findings identify some of the most basic cellular building blocks by which linguistic information is encoded in humans and begin to define the cortical landscape of language at a combined micro (cellular), meso (local population) and macro (regional) scale. Wide-scale recordings reveal neurons in the human brain that encode fundamental components of language such as the grammatical relationships between words, their parts of speech and the higher-order syntactic structure of phrases and sentences.

20.
arXiv (CS.CL) 2026-06-18

RedactionBench

Large Language Models are increasingly applied to sensitive domains that require redaction of personally identifiable information (PII). While redacting PII is a data cleaning prerequisite, existing benchmarks conflate extraction mechanics with privacy semantics. A public phone number is not equivalent to a phone number in a medical record. Whether information constitutes a violation depends heavily on who holds it, why, and in what context, fundamentally differentiating redaction from simple entity recognition. Grounded in contextual integrity, we introduce RedactionBench, a manually annotated benchmark comprising 200 diverse documents across 11 domains, mostly seeded from real-world sources. We also introduce R-Score, a novel character-level metric that treats semantically similar redactions equally and nullifies shallow formatting choices, such as varying masking styles for phone numbers. Evaluations across Named Entity Recognition models, entity extraction Small Language Models, and frontier models equipped with agentic tools demonstrate that contextual redaction remains an unsolved problem. A human evaluation with over 80 users on RedactionBench reveals a stark dichotomy in privacy perceptions. Annotators show consensus with target labels for mandatory redactions (89.4 percent) and safe text preservations (94.1 percent), but fail to agree on contextual redactions (47.7 percent). This variance demonstrates the subjective nature of contextual privacy and motivates R-Score, which decouples contextual ambiguity from strict precision. We compare 35 models across families and report their performance in redacting PII. Finally, we release RedactionBench to establish a baseline for future privacy-preserving systems, hoping to inspire efficient model design and standardized evaluations.

21.
arXiv (CS.LG) 2026-06-16

Circuit Tracing in Autoregressive Protein Language Models

arXiv:2606.16044v1 Announce Type: new Abstract: Protein language models (pLMs) can generate novel protein sequences with properties beyond those observed in nature, yet the mechanisms underlying protein generation remain poorly understood. Existing mechanistic interpretability methods based on sparse autoencoders and transcoders primarily focus on protein representation learning models and do not capture the computation required for autoregressive generation. Here, we introduce ProGenMech, a mechanistic interpretability framework for generative protein language models that extends cross-layer transcoders (CLTs) to ProGen3, a sparse Mixture-of-Experts model trained for both causal generation and span infilling. Unlike per-layer approaches, CLTs reconstruct each layer using sparse latent variables from all preceding layers, enabling faithful recovery of inter-layer generative computation. We further develop a zero-shot circuit discovery framework to identify sparse latent circuits responsible for protein generation and fitness prediction. In causal generation and zero-shot fitness estimation tasks, ProGenMech outperforms local transcoder baselines in recovering ProGen3's probability distribution and functional scoring behavior, while matching the original model's generative distribution in span infilling tasks. Moreover, the recovered circuits reveal biologically meaningful motifs and functional regions associated with conserved sequence patterns and protein fitness landscapes, establishing a foundation for interpretable and steerable protein generation.

22.
arXiv (CS.AI) 2026-06-19

Too long; didn't solve

arXiv:2604.07593v2 Announce Type: replace Abstract: Mathematical benchmarks consisting of a range of mathematics problems are widely used to evaluate the reasoning abilities of large language models, yet little is known about how their structural properties influence model behaviour. In this work, we investigate two structural length variables, prompt length and solution length, and analyse how they relate to model performance on a newly constructed adversarial dataset of expert-authored mathematics problems. We find that both prompt and solution lengths correlate positively with increased model failure across models. We also include a secondary, exploratory analysis of cross-model disagreement. Under a difficulty-adjusted normalised analysis, both variables retain weak negative associations with realised model separation, slightly stronger for prompt length. Overall, our main robust finding is that structural length is linked to empirical difficulty in this dataset.

23.
arXiv (CS.CL) 2026-06-11

Food4All: An Agentic Framework and Benchmark for Food Resource Navigation with Adaptive User Understanding

Food assistance referral requires conversational agents to translate underspecified, often noisy help-seeking dialogues into locally valid resource recommendations. We present Food4All, an agentic food-resource referral framework and benchmark grounded in 686 structured Indiana food resources. Food4All couples a food-specific search tool with 300 multi-turn evaluation tasks spanning single food needs, composite cases with access or document constraints, and five non-ideal user interaction traits: unreasonable demands, rambling responses, impatience, incomplete answers, and inconsistent information. We evaluate six Large Language Models (LLMs) on requirement grounding, resource retrieval, final referral correctness, and interaction efficiency. Although the strongest model achieves 96.33% referral accuracy, our diagnostics reveal persistent failures in grounding schedule, eligibility, intake, and document constraints, as well as failures to preserve valid retrieved resources in the final recommendation. Trait-level analysis further shows that different non-ideal behaviors stress different parts of the referral pipeline. Food4All provides a controlled testbed for studying tool-calling agents in constraint-sensitive food assistance referral under realistic user interaction challenges.

24.
arXiv (CS.AI) 2026-06-18

WebSP-Eval: Evaluating Web Agents on Website Security and Privacy Tasks

arXiv:2604.06367v2 Announce Type: replace-cross Abstract: Web agents automate browser tasks, ranging from simple form completion to complex workflows like ordering groceries. While current benchmarks evaluate general-purpose performance~(e.g., WebArena) or safety against malicious actions~(e.g., SafeArena), no existing framework assesses an agent's ability to successfully execute user-facing website security and privacy tasks, such as managing cookie preferences, configuring privacy-sensitive account settings, or revoking inactive sessions. To address this gap, we introduce WebSP-Eval, an evaluation framework for measuring web agent performance on website security and privacy tasks. WebSP-Eval comprises 1) a manually crafted task dataset of 200 task instances across 28 websites; 2) a robust agentic system supporting account and initial state management across runs using a custom Google Chrome extension; and 3) an automated evaluator. We evaluate a total of 8 web agent instantiations using state-of-the-art multimodal large language models, conducting a fine-grained analysis across websites, task categories, and UI elements. Our evaluation reveals that current models suffer from limited autonomous exploration capabilities to reliably solve website security and privacy tasks, and struggle with specific task categories and websites. Crucially, we identify stateful UI elements are a primary reason for agent failure, with toggles causing more than 45% task failure across many models.

25.
arXiv (CS.LG) 2026-06-15

Decompose Sparsely Where You Should, Absorb Densely Where You Should No

arXiv:2606.14040v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) are typically trained to reconstruct the entire residual stream through a sparse dictionary, implicitly assuming that all activation content is amenable to sparse, monosemantic decomposition. We question this assumption and hypothesize that activations contain a low-rank, dense component that is computationally important to the model yet inherently unsuitable for sparse representation, which serves as a major source of the persistent dense latents widely observed in trained SAEs. To test this, we add a small rank-$r$ linear bottleneck in parallel with standard SAEs (BatchTopK and Matryoshka), allowing dense structure to be absorbed before sparse reconstruction. On Gemma-2-2B layer 12, a rank-24 bottleneck reduces dense latent count by up to 84\% while improving sparse probing and targeted probe perturbation on both architectures at matched sparsity. The absorbed component is (i) structurally identifiable as the top principal components and outlier dimensions; (ii) causally necessary, with removing it raising next-token cross-entropy by 7.5$\times$, far exceeding the 2.8$\times$ from removing the geometrically near-identical top-24 PCA directions; and (iii) redundantly encoded by sparse dictionaries, with ablating 787 maximally aligned sparse features raising cross-entropy by only 2.9$\times$ and ablating 2,048 topic-aligned features leaving MMLU topic classification virtually unchanged, whereas removing the scaffold drops it from 98.7\% to chance. Together, our findings identify a compact, semantically informative and causally important component of residual stream activations (which we term a computational scaffold) that standard sparse dictionaries represent inefficiently, suggesting that the scope of sparsity-based interpretability methods warrants careful re-examination.